Gain-Scheduled Controller Design

Slovak University of Technology in Bratislava Faculty of Electrical Engineering and Information Technology Institute of Robotics and Cybernetics Doctoral Thesis Gain-Scheduled Controller Design Author: Adrian Ilka, Ing. Supervisor: Vojtech Veselý, Prof. Ing. DrSc. A thesis submitted in fulfilment of the requirements for the degree of Doctor of Philosophy at the Institute of Robotics and Cybernetics May 2015

Slovenská technická univerzita v Bratislave Ústav robotiky a kybernetiky Fakulta elektrotechniky a informatiky ZADANIE DIZERTAČNEJ PRÁCE Autor práce: Ing. Adrian Ilka Študijný program: Kybernetika Študijný odbor: 9.2.7. kybernetika Evidenčné číslo: FEI-10836-51270 ID študenta: 51270 Vedúci práce: Miesto vypracovania: Názov práce: prof. Ing. Vojtech Veselý, DrSc. ÚRK Riadenie systémov metódou gain scheduling Špecifikácia zadania: Špecifikácia zadania: Dizertačná práca bude venovaná problematike návrhu regulátora s plánovaným zosilnením (gain-scheduled). Ciel om práce je nájst systematický postup na návrh optimálnych (suboptimálnych) regulátorov s plánovaným zosilnením pri obmedzení vstupno/výstupných hodnôt systémov. Návrh realizujte aj pre nelineárne systémy s neurčitost ami. Dátum zadania: 24. 08. 2012 Dátum odovzdania: 25. 05. 2015 Ing. Adrian Ilka riešitel prof. Ing. Ján Murgaš, PhD. vedúci pracoviska prof. Ing. Ján Murgaš, PhD. garant študijného programu

You know that children are growing up when they start asking questions that have answers. John J. Plomp

Abstract (English) SLOVAK UNIVERSITY OF TECHNOLOGY IN BRATISLAVA Faculty of Electrical Engineering and Information Technology Institute of Robotics and Cybernetics Doctor of Philosophy Gain-Scheduled Controller Design by Adrian Ilka, Ing. This thesis is devoted to controller synthesis, i.e. finding a systematic procedure to determine the optimal (sub-optimal) controller parameters which guarantee the closed-loop stability and guaranteed cost for uncertain nonlinear systems with considering input/output constraints, all this without on-line optimization. The controller in this thesis is given in a feedback structure that is, the controller has information about the system and uses this information to influence the system. In this thesis the linear parameter-varying based gain scheduling is investigated. The nonlinear system is transformed to a linear parameter-varying system, which is used for controller design, i.e. a gain-scheduled controller design with consideration of the objectives on the system. The gain-scheduled controller synthesis in this thesis is based on the Lyapunov theory of stability as well as on the Bellman-Lyapunov function. Several forms of parameter dependent/quadratic Lyapunov functions are presented and tested. To achieve performance quality, a quadratic cost function and its modifications known from LQ theory are used. In this thesis one can also find an application of gain scheduling in switched and in model predictive control with consideration of input/output constraints. The main results for controller synthesis are in the form of bilinear matrix inequalities (BMI) and/or linear matrix inequalities (LMI). For controller synthesis one can use a free and open source BMI solver PenLab or LMI solvers LMILab or SeDuMi. The synthesis can be done in a computationally tractable and systematic way, therefore the linear parametervarying based gain scheduling approach presented in this thesis is a worthy competitor to other controller synthesis methods for nonlinear systems. Keywords: Gain-scheduled control; Lyapunov theory of stability; Guaranteed cost control; Bellman-Lyapunov function; LPV system; Robust control; Input/output constraints

Abstrakt (Slovak) SLOVENSKÁ TECHNICKÁ UNIVERZITA V BRATISLAVE Fakulta elektrotechniky a informatiky Ústav robotiky a kybernetiky Doctor of Philosophy Riadenie systémov metódou gain scheduling Adrian Ilka, Ing. Táto práca sa venuje problematike návrhu regulátora, tj. nájst systematický postup na návrh optimálnych (suboptimálnych) parametrov regulátora, ktoré garantujú stabilitu a kvalitu v uzavretej slučke, pri obmedzení vstupno-výstupných hodnôt systémov pre nelineárne systémy s neurčitost ami, a to bez on-line optimalizácie. Uvedený regulátor má spätno-väzobnú riadiacu štruktúru, čo znamená, že disponuje informáciami o danom systéme, ktoré využíva k jeho ovplyvneniu. Táto práca sa podrobnejšie zaoberá s riadením s plánovaným zosilnením, a to na báze parametricky závislých lineárnych systémov. Nelineárny systém je pretransformovaný na parametricky závislý lineárny systém, čo sa následne využíva na návrh regulátora, tj. regulátora s plánovaným zosilnením, s ohl adom na požiadavky daného systému. Syntéza regulátora s plánovaným zosilnením sa uskutoční na báze Lyapunovej teórie stability s použitím Bellman- Lyapunovej funkcie, v rámci čoho sú prezentované a testované rôzne typy kvadratickej a parametricky závislej Lyapunovej funkcie. Pre dosiahnutie požadovanej kvality sa používa kvadratická účelová funkcia známa z LQ riadenia, s rôznymi modifikáciami. V tejto práci nájdeme aj aplikáciu riadenia s plánovaným zosilnením v oblasti takzvaného prepínacieho riadenia (switched control), ako aj v rámci prediktívneho riadenia (model predictive control). Hlavné výsledky pre syntézu regulátorov sú v tvare bilineárnych maticových nerovníc (BMI) a/alebo linearných maticových nerovníc (LMI). Na návrh regulátorov môžeme používat bezplatný a open source BMI solver PenLab alebo LMI solvre LMILab a SeDuMi. Uvedené skutočnosti umožnia vykonat syntézu jednoduchým a systematickým spôsobom. Riadenie s plánovaným zosilnením na báze parametricky závislých lineárnych systémov prezentované v tejto práci je vhodným konkurentom vo vzt ahu k iným metódam syntézy regulátorov pre nelineárne systémy. Kl účové slová: Riadenie s plánovaným zosilnením; Lyapunová teória stability; Riadenie s garantovanou kvalitou; Bellman-Lyapunová funkcia; LPV systémy; Robustné riadenie; Vstupné/výstupné obmedzenia

Acknowledgements Firstly I would like to thank Vojtech Veselý for encouragement, help and assistance in my research and study. His generous assistance, inspiration and many long discussions almost twice a week greatly contributed to my success during my research and studies. In this context I would also like to thank Alena Kozáková and Danica Rosinová for patient help and many conversations, discussions which helped me to develop some of the results. In particular I am grateful to Boris Rudolf for many long discussions which helped me to clear up several points of mathematical derivations. I would also like to thank Juraj Breza for proofreading the papers and for language corrections. I would like to thank the colleagues at our institute for help and assistance in my research, especially Marian Tárnik for proofreading the papers and letters, Daniel Vozák for materials and help in the course teaching, Štefan Bucz, Ivan Holič, Jozef Škultéty, Ivan Ottinger and Ludwig Tomas for interesting discussions in control, nature and life in general, Beáta Hochschornerová, Alena Foltinová and Jozef Turčánik for making the administrative things run so smoothly, and to all the rest of the staff at the Institute of Robotics and Cybernetics for contributing to the atmosphere. Anna Koláriková and Danka Kurucová from the office for student services for their help during my studies at the university. My gratitude also goes to Michal Kocvara for the open source BMI solver PenLab, to Johan Löfberg for free YALMIP modelling language. I would like to acknowledge thousands of individuals who have coded the L A TEX project for free, mainly Steven Gunn for the L A TEX template used for this thesis. It is due to their efforts that we can generate professionally typeset PDFs now. Finally, I would like to thank my wife Viktória Ilková, my family and friends for being so understanding and inspiring.

Contents Abstract v Abstract (Slovak) vii Acknowledgements ix Contents List of Figures and Tables Abbreviations x xv xviii 1 Introduction 1 1.1 Goals & Objectives............................... 2 1.2 Outline..................................... 3 2 Preliminary chapter 5 2.1 Linear parameter-varying systems....................... 5 2.1.1 Introduction to LPV systems..................... 5 2.1.1.1 Application of the LPV systems.............. 6 2.1.2 Stability analysis............................ 7 2.1.2.1 Time variations and instability............... 7 2.1.2.2 Slow time parameter variations............... 8 2.1.2.3 Arbitrary time parameter variations............ 10 2.1.3 Summary................................ 13 2.2 Gain scheduling................................. 13 2.2.1 Introduction to gain scheduling.................... 14 2.2.1.1 History of gain scheduling.................. 14 2.2.1.2 Application of gain scheduling............... 15 2.2.2 Classical gain scheduling........................ 16 2.2.3 LFT and LPV based gain scheduling................. 17 2.2.3.1 LPV gain scheduling..................... 18 2.2.3.2 LFT gain scheduling..................... 19 2.2.4 Fuzzy gain scheduling......................... 20 2.2.5 Summary................................ 21 xi

Contents 2.3 Discussion.................................... 22 3 Summary of included papers 31 3.1 Introduction................................... 31 3.2 Summary of included papers......................... 32 4 Gain-scheduled PID controller design (Paper 1) 37 4.1 Introduction................................... 37 4.2 Preliminaries and problem formulation.................... 40 4.3 Main results................................... 41 4.4 Examples.................................... 45 4.5 Conclusion................................... 50 5 Gain-Scheduled Controller Design: Variable Weighting Approach (Paper 2) 53 5.1 Introduction................................... 53 5.2 Preliminaries and Problem Formulation................... 54 5.3 Main Results.................................. 56 5.4 Example..................................... 58 5.5 Conclusion................................... 62 6 Design of Robust Gain-Scheduled PI Controllers (Paper 3) 63 6.1 Introduction................................... 63 6.2 Problem formulation and preliminaries.................... 65 6.3 Main results................................... 67 6.4 LMI gain-scheduled robust controller design................. 70 6.4.1 Equivalent gain-scheduled system.................. 71 6.4.2 LMI robust controller design procedure............... 72 6.5 Examples.................................... 74 6.6 Conclusion................................... 80 6.7 Appendix.................................... 81 6.7.1 Linearization of (6.32)......................... 81 6.7.2 Linearization of (6.31)......................... 81 7 Robust Gain-scheduled PID Controller Design for uncertain LPV systems (Paper 4) 85 7.1 Introduction................................... 85 7.2 Problem formulation and preliminaries.................... 87 7.3 Main Results.................................. 90 7.4 Examples.................................... 93 7.5 Conclusion................................... 97 8 Robust Controller Design for T1DM Individualized Model: Gain Scheduling Approach (Paper 5) 101 8.1 Introduction................................... 101 8.2 Problem formulation and preliminaries.................... 102 8.2.1 T1DM model.............................. 103 8.2.2 Identification of model parameters.................. 104 xii

Contents 8.2.2.1 Insulin absorption subsystem:............... 104 8.2.2.2 Insulin sensitivity index and insulin action time:..... 104 8.2.2.3 Finalizing the model:.................... 105 8.3 LPV-based robust gain-scheduled controller design............. 105 8.3.1 LPV model of T1DM......................... 105 8.3.2 Robust gain-scheduled controller design............... 107 8.4 Simulation experiments............................ 111 8.5 Conclusion................................... 111 9 Novel approach to switched controller design for linear continuous-time systems (Paper 6) 115 9.1 Introduction................................... 115 9.2 Preliminaries and problem formulation.................... 117 9.3 Switched controller design........................... 119 9.3.1 Quadratic stability approach..................... 120 9.3.2 Multiple Lyapunov function approach................ 121 9.4 Examples.................................... 124 9.5 Conclusion................................... 131 10 Robust Switched Controller Design for Nonlinear Continuous Systems (Paper 7) 133 10.1 Introduction................................... 133 10.2 Problem statement and preliminaries..................... 135 10.2.1 Uncertain LPV plant model for switched systems.......... 135 10.2.2 Problem formulation.......................... 136 10.3 Main results................................... 137 10.4 Example..................................... 141 10.5 Conclusion................................... 143 11 Gain-Scheduled MPC Design for Nonlinear Systems with Input Constraints (Paper 8) 147 11.1 Introduction................................... 147 11.2 Problem formulation and preliminaries.................... 149 11.2.1 Case of finite prediction horizon................... 149 11.2.2 Case of infinite prediction horizon.................. 152 11.3 Main results................................... 154 11.3.1 Finite prediction horizon........................ 154 11.3.2 Infinite prediction horizon....................... 155 11.4 Examples.................................... 156 11.5 Conclusion................................... 159 12 Unified Robust Gain-Scheduled and Switched Controller Design for Linear Continuous-Time Systems (Paper 9) 163 12.1 Introduction................................... 163 12.2 Problem formulation and preliminaries.................... 165 12.3 Robust gain-scheduled and switched controller design............ 167 12.4 Examples.................................... 170 12.4.1 Prescribed controller structure.................... 170 xiii

Contents 12.4.2 Different quadratic stability approach................ 171 12.4.3 Robust controller design........................ 171 12.5 Conclusion................................... 176 13 Concluding remarks 179 13.1 Brief overview.................................. 179 13.2 Closing remarks and future works....................... 179 A List of publications 181 xiv

List of Figures and Tables Figures 2.1 Instability induced by switching dynamics.................. 8 2.2 Stability and peaking.............................. 9 2.3 Integral for Lyapunov function construction................. 11 2.4 The time-line of gain scheduling........................ 15 2.5 LFT M structure............................. 19 4.1 Exogenous signal α(t)............................. 45 4.2 Simulation results............................... 47 4.3 θ(t), α(t)..................................... 47 4.4 Simulation results for θ = 1.......................... 48 4.5 Simulation results for θ = 0.......................... 48 4.6 Simulation results for θ = 1......................... 48 4.7 Simulation results for θ 1, 1....................... 49 4.8 Simulation results for R = 1, Q = 1 10 1, S = 1 10 3........... 51 5.1 Exogenous signal α(t)............................. 58 5.2 Simulation results with GSC (5.15)...................... 60 5.3 Simulation results with GSC (5.15) zoomed................ 60 5.4 Simulation results with GSC (5.16)...................... 60 5.5 Simulation results with GSC (5.16) zoomed................ 60 5.6 Simulation results with GSC (5.17)...................... 61 5.7 Simulation results with GSC (5.17) zoomed................ 61 5.8 Simulation results (y(t), w(t), u(t)) with GSC (5.16)............ 61 5.9 Simulation results (θ(t), α(t)) with GSC (5.16)............... 61 6.1 λ θ (ω) at θ 1 = θ 1 sin ωt............................. 76 6.2 λ ξ (ω) at θ 1 = 0, A no (1, 1) = 1.25 + 0.25 sin ωt............... 76 6.3 λ θ (ω) at θ 1 = θ 1 sin ωt............................. 77 6.4 Dynamic behaviour of the closed-loop system for case DP 1........ 78 6.5 λ θ (ω) at θ 1 = θ 1 sin ωt, θ 2 = θ 2 sin(ωt + 90)................. 79 6.6 Dynamic behaviour of closed-loop system for case DP 2.......... 79 6.7 Dynamic behaviour of closed-loop system.................. 80 7.1 Simulation results at QS3, γ = 1, α 0, 100................ 95 7.2 Simulation results at QS2, γ = 1, α 0, 100................ 95 7.3 Simulation results at QS1, γ = 1, θ 1 0, 1................. 97 xv

List of Figures and Tables 7.4 Simulation results at θ 1 = 0.......................... 97 8.1 Bergman s model with the discrete LPV model............... 111 8.2 Simulation results for time period of 4 days................. 112 9.1 Simulation results w(t), y(t) with switched controller (9.31) QS..... 125 9.2 Simulation results w(t), y(t) with switched controller (9.32) MPQS... 126 9.3 Calculated switched parameters θ(t) and the switching signal α(t) for the case of θ i = 1000 [1/s]............................. 126 9.4 Simulation results w(t), y(t) with switched controller (9.33) MPQS... 127 9.5 Calculated switching parameters θ(t) and the switching signal α(t) for the case of θ i = 2000 [1/s]........................... 127 9.6 Simulation results w(t), y(t) with switched controller (9.37) MPQS... 128 9.7 Calculated switching parameters θ(t) and the controller output with switched controller (9.37) MPQS........................... 128 9.8 Simulation results w(t), y(t) with switched controller (9.39)........ 130 9.9 Switched Controller output (9.39)....................... 130 9.10 Time delay changes............................... 130 10.1 Simulation results w(t), y(t) with the first gain-scheduled switched controller...................................... 143 10.2 Simulation results w(t), y(t) with second gain-scheduled switched controller........................................ 143 10.3 Simulation results w(t), y(t) with the first gain-scheduled switched controller zoomed................................ 143 10.4 Simulation results w(t), y(t) with second gain-scheduled switched controller zoomed.................................. 143 10.5 Development of the switching variable α 1 (t)................. 144 10.6 Development of the switching variable α 1 (t)................. 144 11.1 System output y(t) and the setpoint w(t).................. 157 11.2 System input, u(t)............................... 158 11.3 Calculated scheduling parameters, θ(t).................... 158 11.4 Simulation results, y(t), w(t).......................... 159 11.5 System input, u(t)............................... 159 11.6 Calculated scheduling parameters, θ(t).................... 159 12.1 Simulation results y(t), w(t) for the first case of robust gain-scheduled controllers.................................... 175 12.2 Calculated scheduled parameters θ DP 1 4 (t) for the first case of robust gain-scheduled controllers........................... 175 12.3 Simulation results y(t), w(t) for the second case of robust gain-scheduled controllers.................................... 175 12.4 Scheduled parameter θ 1 (t) for the second case of robust gain-scheduled controllers.................................... 175 xvi

List of Figures and Tables Tables 8.1 T1DM identified model parameters...................... 105 8.2 Other fixed parameters of the model..................... 105 xvii

Abbreviations A/D AQS BMI BW CGM DPC GS GSC GSLMIDP LFT LMI LPV LQ LTI MIMO MPC MPQS MU NCSs PDQS PID PK SISO T1DM WP Analog / Digital Affine Quadratic Sstability Bilinear Matrix Inequality Body Weight Continuous Glucose Monitoring Dynamic Property Coefficient Gain-Scheduling Gain-Scheduled Controller Gain-Scheduled Linear Matrix Inequality Design Procedure Linear Fractional Transformation Linear Matrix Inequality Linear Parameter-Varying Linear-Quadratic Linear Time Invariant Multi Input Multi Output Model Predictive Control Multi Parameter Quadratic Stability Machine Unit Networked Control Systems Parameter Dependent Quadratic Stability Proportional-Integral-Derivative PharmacoKinetics Single Input Single Output Type One Diabetes Mellitus Working Point xix

To my wonderful wife Viktória

1 Introduction This thesis is devoted to controller synthesis, i.e. finding a systematic procedure to determine the optimal (sub-optimal) controller parameters which guarantee the closedloop stability and guaranteed cost for uncertain nonlinear systems with considering input/output constraints. In consideration of the objectives stated for the system such as tracking a reference signal or keeping the plant at a desired working point (operation point) and based on the knowledge of the system (plant), the controller takes decisions. In this thesis, the controller is given in a feedback structure, which means that the controller has information about the system and uses it to influence the system. A system with a feedback controller is said to be a closed-loop system. To design a controller which satisfies the objectives, we need an adequately accurate model of the physical system. Nevertheless, real plants are hard to describe exactly. Alternatively, the designed controller must handle cases when the state of the real plant differs from what is observed by the model. A controller that is able to handle model uncertainties and/or disturbances is said to be robust, and the theory dealing with these issues is said to be robust control. The robust control theory is well established for linear systems but almost all real processes are more or less nonlinear. If the plant operating region is small, one can use robust control approaches to design a linear robust controller, where the nonlinearities are treated as model uncertainties. However, for real nonlinear processes, where the operating region is large, the above mentioned controller synthesis may be inapplicable because the linear robust controller may not be able to meet the performance specifications. For this reason, the controller design for nonlinear systems is nowadays a very determinative and important field of research. Gain scheduling is one of the most commonly used controller design approaches for nonlinear systems and has a wide range of use in industrial applications. Many of the early articles were associated with flight control and aerospace. Then, gradually, this approach has been used almost everywhere in control engineering, which was greatly advanced with the introduction of LPV systems. 1

1. Introduction Linear parameter-varying systems are time-varying plants whose state space matrices are fixed functions of some vector of varying parameters θ(t). These were introduced first by Jeff S. Shamma in 1988 to model gain scheduling. Today the LPV paradigm has become a standard formalism in the area of systems and controls with lot of contributions and articles devoted to analysis, controller design and system identification of these models. This thesis deals with linear parameter-varying based gain scheduling, which means that the nonlinear system is transformed to a linear parameter-varying system, which is used to design a controller, i.e. a gain-scheduled controller. The problem formulation is close to the linear system counterpart, therefore using LPV models for controller design has potential computational advantages over other controller synthesis methods for nonlinear systems. Not to mention that the LPV based gain scheduling approaches comes with a theoretical validity because the closed-loop system can meet certain specifications. Nonetheless, following the literature it is ascertainable that there are still many unsolved problems. This thesis is devoted to some of these problems. 1.1 Goals & Objectives As already mentioned, there are many unsolved problems. Therefore, it is necessary to find new and novel controller design approaches. The main goal of this thesis is to find a controller design approach for uncertain nonlinear systems, which guarantees the closed-loop stability and the optimal controller output with considering input/output constraints, all this without on-line optimization and need of high-performance industrial computers. In order to achieve the above mentioned goal, we have set the following objectives: To suggest a gain-scheduled PID controller design approach with guaranteed cost in continuous and discrete time state space using BMI To suggest a robust gain-scheduled PID controller design approach with guaranteed cost and parameter dependent quadratic stability in state space using BMI To suggest a variable weighting gain-scheduled approach To convert some BMI controller design approaches to LMI To suggest a switched and model predictive gain-scheduled method To suggest a gain-scheduled controller design approach with input/output constraints To apply methods to relevant processes 2

1.2. Outline 1.2 Outline The sequel of this thesis is organized as follows. In the preliminary chapter (Chapter 2), one can find a literature review with a brief overview of linear parameter-varying systems and gain scheduling. Chapter 3 presents an overview of research results with a brief summary of included papers. After this, one can find 9 papers, which cover the main research results obtained within the last 2.5 years (Chapter 4-12). Finally, in Chapter 13, following the papers, some concluding remarks and suggestions for future research are given. 3

2 Preliminary chapter In this chapter preliminaries of linear parameter-varying systems as well as gain scheduling are introduced. This chapter is intended to highlight the properties and give a short background to the tools used in the appended papers. 2.1 Linear parameter-varying systems Linear parameter-varying systems are time-varying plants whose state space matrices are fixed functions of some vector of varying parameters θ(t). It was introduced first by Jeff S. Shamma in 1988 [1] to model gain scheduling. Today LPV paradigm has become a standard formalism in systems and controls with lot of researches and articles devoted to analysis, controller design and system identification of these models, as Shamma wrote in [2]. This section deals with LPV models and presents analytical approaches for LPV systems. 2.1.1 Introduction to LPV systems Linear parameter-varying systems are time-varying plants whose state space matrices are fixed functions of some vector of varying parameters θ(t). Linear parameter-varying (LPV) systems have the following interpretations: they can be viewed as linear time invariant (LTI) plants subject to time-varying known parameters θ(t) θ θ, they can be models of linear time-varying plants, they can be LTI plant models resulting from linearization of the nonlinear plants along trajectories of the parameter θ(t) θ θ which can be measured. 5

2. Preliminary chapter For the first and third class of systems, parameter θ can be exploited for the control strategy to increase the performance of closed-loop systems. Hence, in this thesis the following LPV system will be used: ẋ = A(θ(t))x + B(θ(t))u y = Cx (2.1) where for the affine case A(θ(t)) = A 0 + A 1 θ 1 (t) +... + A p θ p (t) B(θ(t)) = B 0 + B 1 θ 1 (t) +... + B p θ p (t) and x R n is the state, u R m is a control input, y R l is the measurement output vector, A 0, B 0, A i, B i, i = 1, 2..., p, C are constant matrices of appropriate dimension, θ(t) θ θ Ω and θ(t) θ θ Ω t are vectors of time-varying plant parameters which belong to the known boundaries. The LPV paradigm was introduced by Jeff. S. Shamma in his Ph.D. thesis [1] for the analysis of gain-scheduled controller design. The authors in early works (see [1, 3 8] and surveys [9, 10]) in gain scheduling the LPV system framework called as the golden mean between linear and nonlinear dynamics, because the LPV system is an indexed collection of linear systems, in which the indexing parameter is exogenous, i.e., independent of the state. (wrote J. S. Shamma in his Ph.D. thesis [1]). In gain scheduling, this parameter is often a function of the state, and hence endogenous ẋ = A(z)x + B(z)u y = C(z)x z = h(x) (2.2) 2.1.1.1 Application of the LPV systems Since the first publication devoted to LPV systems, the LPV paradigm has been used in several fields in control engineering including the modeling and control design. Traditionally the gain scheduling was the primary design approach for flight control and consequently many of the first articles and papers which applied and improved the LPV framework were associated with flight control. Afterwards continuously many papers and articles have appeared which are using LPV paradigm in several application areas such as: Flight control and missile autopilots [11 17] Aeroelasticity [18 21] Magnetic bearings [22 25] 6

2.1. Linear parameter-varying systems Automotive bearings [26 28] Energy and power systems [29 34] Turbofan engines [35 38] Microgravity [39 41] Diabetes control [42 44] Anesthesia delivery [45] IC manufacturing [46] etc. Due to the success of LPV paradigm in 2012 for the twentieth anniversary of the invention of LPV paradigm a gift edition book was published by Javad Mohammadpour and Carsten W. Scherer Editors at Springer [2] which is fully devoted to LPV systems. 2.1.2 Stability analysis The basic stability analysis question for LPV systems is how to ensure the stability of the closed-loop nonlinear system and of the closed-loop family of linear systems, when the scheduled parameters are changed. The following section is devoted to this basic stability question and shows the basic theoretical approaches to investigate the stability for 1. slow time parameter variations, 2. arbitrarily fast time parameter variations. 2.1.2.1 Time variations and instability It is a well-known problem from linear system analysis that time variations can induce instability. For example, consider a stable LTV system (2.3), so the eigenvalues of A(t) are in the left half plane for all t 0. The question is for which solution the state x(t) grows exponentially. ẋ = A(t)x (2.3) Fig. 2.1 shows the main insight into this problem using the state trajectories of the LPV system (2.4) with parameter θ which is periodically switching between two values θ(t) ω a, ω b. In this figure the red line indicates the unstable switching trajectory and the dashed lines indicate individual oscillatory trajectories. ẋ = ( 0 1 θ 2 0 7 ) x (2.4)

2. Preliminary chapter For a fixed value of θ the LTI system is marginally stable. alignment of phases of increasing magnitude. Instability occurs by an x 2 x 1 Figure 2.1: Instability induced by switching dynamics Concerning to induced instability non-minimum phasedness is induced. The right-halfplane zeros in the transfer function of an LTI system can cause radical limitations in achievable performance. While time-varying systems do not have right-half-plane zeros, there are similar notions and similar resulting limits of performance. Shamma [2] defines a non-minimum phased property for nonlinear time-varying systems, where an unbounded input produces a bounded output. This property produces fundamental limitations on the closed-loop disturbance rejection. As Shmamma presented in [1], parameter time variation can induce instability, they can also induce such non-minimum phased behaviours. Summarizing all of this, an LPV system can be the minimum phase for constant parameter values, but non-minimum phase under time variation and thereby have fundamental limits on achievable performance that are not apparent from the constant parameter analysis. 2.1.2.2 Slow time parameter variations In [2], Shamma has stated the following: Stability for constant parameter-varying parameter trajectories implies stability for slowly time-varying parameter trajectories. This section presents a collection of results which motivated Shamma to formalize the previous statement. Let Θ denote the set of admissible parameter values whereas Q denotes admissible trajectories for θ( ), the related Θ denotes admissible values of θ(t). Let assume that for any θ 0, the LTI system is exponentially stable. According to Shamma, in particular, let m 1 and λ > 0 be such that for any θ 0 Θ, solution of (2.3) satisfy x(t) me λt x(0) where m is referred to as a peaking constant which reflects that the state may increase in magnitude before decaying exponentially. Fig. 2.2 shows the main principle of stability under slow time variations, where the red line indicates the actual state magnitude, the 8

2.1. Linear parameter-varying systems blue line indicates a succession of upper bounds implied by me λt and the green line is an exponentially decaying overall upper bound. For more details see [47]. 5 4 Amplitude 3 2 1 0 0 5 10 15 20 25 30 35 40 45 50 t[s] Figure 2.2: Stability and peaking The statement Slow time-varying for the continuous case can be characterized as follows: Assume Lipschitz continuity of A( ) for some L A > 0 A(θ) A(θ ) L A θ θ (2.5) for all θ, θ θ c. The expression x denotes the Euclidean norm of x R n and A denotes the induced matrix norm. Than Persistently slow: θ < ɛ Slow on average: over any interval [t 0, t 0 + T ] is small. inf sup 1 t0 +T T >0 t 0 0 T θ dt < ɛ t 0 Theorem 2.1. For all of the above settings the LPV system (2.1) is exponentially stable for a sufficiently small ɛ > 0. Stability results for properly slow time variations, trace back to classical results in ordinary differential equations [48]. Nevertheless, a suitable analysis can derive revealing explicit bounds in the above case Persistently slow [1] and slow on average [49]: ɛ < λ 2 4L A m log(m) 9

2. Preliminary chapter Shamma stated an interesting implication from the above bounds, time variations can be arbitrarily fast, when m = 1. In terms of the previous discussion, m = 1 implies that trajectories in the constant parameter case have no peaking, and therefore cannot align to produce instability. 2.1.2.3 Arbitrary time parameter variations This section deals with the stability question from the other extreme, when time variations are arbitrarily fast. For this discussion consider an LPV plant in the form Shamma and others [50 52] concluded that ẋ(t) = A(θ(t))x(t) (2.6) Determining whether solutions of (2.6) are bounded is undecidable. Determining whether (2.6) is asymptotically stable is NP-hard 1 Consequentially, deriving efficient algorithms for assessing stability will remain to be elusive. Shamma according that a consequence of the complexity results is, that one must settle for non-definitive methods or inefficient algorithms to access stability. Theorem 2.2. The LPV system (2.1) is exponentially stable for all θ Ω if there exist symmetric, positive defined matrix P such that the following inequality holds A T (θ)p + P A(θ) < 0 (2.7) The proof is, that x T P x is a Lyapunov function for the LPV system, which in this case is the only Lyapunov function for all associated constant parameter LTI system. The result is only a sufficient condition. Shamma in [2] introduced a simple example from [53] which explains this problem. Consider a simple second-order system whose dynamics matrix can switch between two matrices ẋ {A 1 x, A 2 x} (2.8) This can be viewed as an LPV plant with θ 1, 1. In [53] the authors shows that for ( ) ( ) 1 1 1 a A 1 =, A 2 = 1 1 1/a 1 where 3+ 8 < a < 10, the above system is stable for arbitrary switching, but no P exists satisfying conditions (2.7). This is called by some authors [54 57] as the conservativeness 1 NP-hard (Non-deterministic Polynomial-time hard), in computational complexity theory, is a class of problems that are, informally, at least as hard as the hardest problems in NP 10

2.1. Linear parameter-varying systems of the quadratic stability. Therefore a lot of people were looking for a solution on how to reduce the conservatism. And this has resulted in certain special structures of suitable Lyapunov functions [54 60]. Let denote A(t, τ; θ([τ, t])) (2.9) as the state transition matrix for an LPV system, where the dependence on the parameter trajectory is explicit (over the interval [τ, t]). Accordingly x(t) = A(t, τ; θ([τ, t]))x(τ) (2.10) Assuming that an LPV plant is exponentially stable for all parameter trajectories, there exist m and λ > 0 such that A(t, τ; θ([τ, t])) me λ(t τ) (2.11) Let T be such that me λt < 1, and define the following Lyapunov function candidate (e.g., [61]) V (x, t) = t+t t A(τ, t; θ([t, τ]))x 2 dτ (2.12) Fig. 2.3 illustrates the construction of this function. This function is the energy of the solution over the interval [t, t + T ]. One can show that V (x(t), t) is decreasing along solutions of the LPV system. In particular ( ) V (x(t + h), t + h) V (x(t), t) h x(t) 2 + h x(t + T ) 2 < 1 me λt x 2 (2.13) Neglecting issues of differentiability, the above construction suggests that dv (x(t), t) dt < c x 2 (2.14) 5 4 Amplitude 3 2 1 0 0 5 10 15 20 25 30 35 40 45 50 t t+t t[s] Figure 2.3: Integral for Lyapunov function construction 11

2. Preliminary chapter The structure of this Lyapunov function can be rewritten as a quadratic function in x, where the defining matrix is a function of the future parameter trajectory V (x(t), t) = x T (t)p (θ([t, t + T ]))x(t) (2.15) We can reparametrize the function to be a function of past parameter trajectories V (x(t), t) = x T (t) P (θ([t T, t]))x(t) (2.16) The authors in [59] used a similar construction to derive the following theorem Theorem 2.3. [59] An LPV system is exponentially stable for arbitrary time variations if and only if there exists a trajectory dependent quadratic Lyapunov function of the form V (x, t) = x T P (θ([t T, t]))x (2.17) In discrete time, authors [59] use this result to derive a numerical search for Lyapunov functions. Regarding the previous discussion on complexity, this search may need to admit progressively longer intervals of trajectory dependence. It turns out that one can eliminate the dependence on the parameter trajectory altogether. The intuition is as follows. From the Lyapunov function in Theorem 2.3, define V (x) = inf θ([t T,t]) xt P (θ([t T, t]))x (2.18) The new Lyapunov function is the former Lyapunov function evaluated at a worst case trajectory [2]. Again, an informal analysis illustrates that this parameter-independent Lyapunov function decreases along the solution of the LPV system for all parameter trajectories. This motivates the existence in general of a pseudo-quadratic Lyapunov function. Authors Molchanov and Pyatnitskiy in [60] introduced the following theorem Theorem 2.4. [60] An LPV system is exponentially stable for arbitrary time variations if and only if there exists a Lyapunov function of the form V (x) = x T P (x)x (2.19) for some family of matrices P ( ), with the property that P (αx) = αp (x) for α 0. In book [2], papers [62, 63], survey [64] and monograph [65] we can find further discussions which go on to characterize alternative piecewise linear structures for exponentially stable LPV systems. Besides these we can find papers which investigate the stability of an LPV systems using parameter-dependent quadratic stability [54, 56, 57, 66, 67]. The main principle of parameter-dependent quadratic stability is that against the result with quadratic stability we have one Lyapunov function for all vertices of θ. So the Lyapunov function is parameter-dependent V (θ(t)) = x T (t)p (θ(t))x(t) (2.20) 12

2.2. Gain scheduling where θ Ω and p P (θ) = P 0 + P i θ i (2.21) i=1 P. Gahinet, P. Apkarian and M. Chilali in [57] in this context introduced the following theorem Theorem 2.5. [57] The LPV system (2.1) for θ Ω and θ Ω t is affinely quadratically stable if and only if there exist p + 1 symmetric matrices P 0, P 1,..., P p such that P (θ) = P 0 + p P i θ i > 0 (2.22) and for the first derivative of Lyapunov function V (θ) = x T P (θ)x along the trajectory of LPV system (2.1) it holds dv (x, θ) dt i=1 = x T ( A(θ) T P (θ) + P (θ)a(θ) + ) dp (θ) x < 0 (2.23) dt where dp (θ) dt = p P i θi i=1 p P i ρ i i=1 In this case we must have predefined the maximum rate of change of scheduled parameters θ i as ρ i. 2.1.3 Summary In this section (Linear parameter-varying system) the LPV systems were presented and described and their stability analysis since its introduction (1988 by Jeff. S. Shamma [1]) to the present (2015). The analysis and theorems stated herein are presented in an informal manner. Technical details may (and should) be found in the associated references. 2.2 Gain scheduling The robust control theory is well established for linear systems but almost all real processes are more or less nonlinear. If the plant operating region is small, one can use the robust control approaches to design a linear robust controller where the nonlinearities are treated as model uncertainties. However, for real nonlinear processes, where the operating region is large, the above mentioned controller synthesis may be inapplicable. For this reason the controller design for nonlinear systems is nowadays a very determinative and important field of research. 13

2. Preliminary chapter Gain scheduling is one of the most common used controller design approaches for nonlinear systems and has a wide range of use in industrial applications. In this section the main principles, several classical approaches and finally the linear parameter-varying based version of gain scheduling are presented and investigated. 2.2.1 Introduction to gain scheduling In literature a lot of term are meant under gain scheduling (GS). For example switching or blending of gain values of controllers or models, switching or blending of complete controllers or models or adapt (schedule) controller parameters or model parameters according to different operating conditions. A common feature is the sense of decomposing nonlinear design problems into linear or nonlinear sub-problems. The main difference lies in the realization. Consequently gain scheduling may be classified in different way According to decomposition 1. GS methods decomposing nonlinear design problems into linear sub-problems 2. GS methods decomposing nonlinear design problems into nonlinear (affine) sub-problems According to signal processing 1. Continuous gain scheduling methods 2. Discrete gain scheduling methods 3. hybrid or switched gain scheduling methods According to main approaches 1. Classical (linearization based) gain scheduling 2. LFT based GS synthesis 3. LPV based GS synthesis 4. Fuzzy GS techniques 5. Other modern GS techniques 2.2.1.1 History of gain scheduling The ferret in the history of gain scheduling appears in the 1960s but a similar simpler technique was used in World War II toat control the rockets V2 (switching controllers based on measured data). It is not surprising therefore that gain scheduling as a concept or notion firstly appear in flight control and later in aerospace. Leith and Leithead in their survey [9] and likewise also Rugh and Shamma in their survey paper [10] considered the first appearance of GS from the 1960s. Rugh stated in his survey that Gain control 14

2.2. Gain scheduling does appear in the 25th Anniversary Index (1956 1981) published in 1981 but only one of the five listed papers is relevant to gain scheduling. Also Automatica lists gain scheduling as a subject in its 1963 1995 cumulative index published in 1995. Of the four citations given, only one dated earlier than 1990 [1]. It can be stated that increased attention to gain scheduling appeared after introducing the LPV paradigm by Jeff. S. Shamma (1988). This is partly understandable because LPV paradigm allowed to describe nonlinear system as a family of linear systems and hence investigate the stability of these systems. Fig. 2.4 shows the major dates with remarks in a time-line of gain scheduling. Quiet years; only few publications devoted to gain-scheduling Jeff. S. Shamma introduced LPV systmes Gain-scheduling is one of the most popular approaches to nonlinear control design Past 1991 Present 1943 1960 1969 1981 1988 2010 Fist gain-scheduling like controllers; II. World War First appear of notion gainscheduling; application in flight and aerospace "Gain control" does appear in the 25th Anniversary Index Rugh and Shamma and also Leith and Leithead survay papers on gainscheduling; Increased interest in gain-scheduling Figure 2.4: The time-line of gain scheduling 2.2.1.2 Application of gain scheduling As already noted, traditionally the gain scheduling was the primary design approach to flight control and, consequently, many of the first articles and papers were associated with flight control [68 75] and aerospace [76 78]. Then gradually GS has been used almost everywhere in control engineering which was greatly advanced with the introduction of LPV systems. The second big bang in the history of gain scheduling was the advent of fuzzy gain scheduling. Today, every second paper that appears under gain scheduling is devoted to fuzzy gain scheduling. Due to this wide range of gain-schedule approaches, gain scheduling is now used in several fields in practice. For example in power systems the gain scheduling enjoyed exceptional success in control of wind turbines [79 85]. But beside all this, some papers are devoted to hydro turbines [86, 87], gas turbines [88], power system stabilizers [89] and generators [90]. Many papers in gain scheduling are devoted to magnetic bearings [91 96] but we can find some papers devoted to also to microgravity [97], turbofan engine [98] and diabetes control [99]. 15

2. Preliminary chapter 2.2.2 Classical gain scheduling In the case of nonlinear dynamics an idea is widely used among control engineers to linearize the plant around several operating points and to use linear control tools to design a controller for each of these points. The actual controller is implemented using the gain scheduling approach. Success of such an approach depends on establishing the relationship between a nonlinear system and a family of linear ones. There are two main problems: 1. Stability results: stability of the closed-loop nonlinear system and of the closedloop family of linear systems, when scheduled parameters are changes. 2. Approximation results which provide a direct relationship between the solution of closed-loop nonlinear systems and the solution of associated linear systems [10], [9] Rugh and Shamma in [10] comprise four main steps in classical gain scheduling 1. A family of LTI approximations are obtained from nonlinear plant at constant operating points (equilibria), parametrized by exogenous signal θ (scheduled parameter) which is computed using linearization based scheduling. The linearization has to correspond to zero error. Other syntheses to derive a parameter-dependent model are Off-equilibrium or velocity based linearization [9, 100 102] - when zero equilibrium points or working conditions are not present Quasi LPV approach [9, 10, 102], in which the plant dynamics are rewritten to distinguish nonlinearities as time-varying parameters that are used as scheduling variables. Direct LPV modelling, based on a linear plant incorporating time-varying parameters [1, 75, 102] - when no nonlinear plant is involved. This also includes black-box or data-based modelling methods 2. A set of LTI controllers are designed using linear control tools for previously derived set of local LTI models to achieve specified performance and stability at each operating point. The resulting set of controllers is also parametrized by scheduled parameter θ. Although the scheduled parameter is time-varying, the classical gain scheduling methods are based on fixed or frozen scheduling parameter values. To enable subsequent scheduling, interpolation of controller parameters, the set of LTI controllers almost requires a fixed structure of the controller design. Exceptions are in the case direct derivation of a Linear Parameter-Varying (LPV) controller for a corresponding LPV plant model is possible, subsequent scheduling, interpolation becomes superfluous. 16

2.2. Gain scheduling when discrete or hybrid scheduling instead of continuous scheduling is demanded, the set of controller designs not necessarily need to be fixed-structured. 3. Implementation of the family of LTI controllers such that the controller coefficients are scheduled according to the current value of the scheduling variable, e.g. by controller gain interpolation or scheduling. At this point, θ = θ(t) is implemented. At each operating point, the scheduled controller has to be linearized to the corresponding linear controller design as well as provide a constant control value yielding zero error at these points. As mentioned in Step 2, in the case of direct scheduling, this step becomes superfluous. Furthermore, in the case of discrete scheduling, the implementation of the LTI controllers involves the design of a scheduled selection procedure that is applied to the set of LTI controllers, rather than the design of a family of scheduled controllers. The presence of hidden coupling terms is an important aspect which yields various additional requirements to the scheduling procedure. 4. Typically, local performance assessment can be performed analytically, whereas assessment of global performance and robustness has to be established by extensive simulations. Non-local performance of the gain-scheduled controller is evaluated and checked by simulations. 2.2.3 LFT and LPV based gain scheduling The LPV and LFT syntheses are based on LPV and LFT plant representations respectively (Naus [102]). Both methods yield direct synthesis of a controller utilizing (L 2 or H ) norm based methods, with guarantees the robustness, performance and nominal stability of the overall gain-scheduled design [7, 57, 66, 102 104]. LPV and LFT syntheses essentially involve only two main steps. 1. The first step corresponds to the classical approach. A family of LTI approximations of a nonlinear plant at constant operating points (equilibria), parameterized by constant values of convenient plant variables or exogenous signals θ are computed. Subsequent implementation of the controller requires θ = θ(t) to be a measurable variable. Besides the already mentioned methods, which all arrive at Linear Parameter-Varying (LPV) models, in specific cases a LFT description is possible. The LFT description serves as a basis for subsequent LFT controller synthesis. 2. LPV and LFT control synthesis directly yield a gain-scheduled controller. Stability and performance specifications can be guaranteed a priori as the time-varying parameter θ(t) instead of its corresponding frozen value θ is addressed in the design process. In [102] one can find only continuous-time gain scheduling but the author Sato in 2011 [105] introduced discrete-time version of LPV based gain scheduling where stability investigated with both H 2 and H. 17